library(tensorflow)library(EBImage)library(XML)library(stats)library(RJSONIO)library(rPython)library(leaflet)library(htmlwidgets)python.load("omrekenen.py")We want to determine the location of buoys using a series of gps labeled photos.This allows us to place cameras on ships that can automatically alarm us, if anything were to be missing.We want to determine the location of buoys using a series of gps labeled photos.
This allows us to place cameras on ships that can automatically alarm us, if anything were to be missing.
In the folder db/test we find three photos. The csv file db/test_location gives us the gps locations of the photos. We display the data herunder.In the folder db/test we find three photos. The csv file db/test_location gives us the gps locations of the photos. We display the data herunder.
photos = list.files('db/test', full.name = TRUE)for(im in photos){im = readImage(im)display(im) }
locations = read.csv('db/test_locations.csv', header = FALSE)colnames(locations) = c('lat', 'lon')print(locations)The first task, and the task in which we preform some deeplearning, is to determine which pixels in the photo belong to the buoy and which pixels do not.For this we train an network to classify the pixels in the image. In the folder db/boeien we have a large number of photos of buoys and in train_labels.csv we have for each photo a square indicating the position of the buoy.Let's have a look at the positions tableThe first task, and the task in which we preform some deeplearning, is to determine which pixels in the photo belong to the buoy and which pixels do not.
For this we train an network to classify the pixels in the image. In the folder db/boeien we have a large number of photos of buoys and in train_labels.csv we have for each photo a square indicating the position of the buoy.
Let's have a look at the positions table
table = read.csv('db/train_labels.csv')head(table)Let's display the image of a random row i of the table and let's draw a square according to the xmax, ymax, xmin and yminLet's display the image of a random row i of the table and let's draw a square according to the xmax, ymax, xmin and ymin
i = sample(x = c(1:nrow(table)), size = 1)im = table$naam[i]im = readImage(paste0( im))display(im)segmentation = array(0, dim = dim(im)[1:2])segmentation[table$xmin[i]:table$xmax[i], table$ymin[i]:table$ymax[i]] = 1display(segmentation)
The data seems to be allright.The data seems to be allright.
Now lets built a network to reproduce such an image segmentation.Now lets built a network to reproduce such an image segmentation.
clas = as.integer(2) #aantal klassenlr = 1e-2 #learningratekanalen = as.integer(3) #kanalen imageh= as.integer(972)w = as.integer(1296)#############################place holders#inputx <- tf$placeholder(tf$float32, shape(NULL, w , h, kanalen))#target valueslabel <- tf$placeholder(tf$int32, shape(NULL,w, h))############variabelenw_conv1 <-tf$Variable(tf$truncated_normal(shape(10L, 10L, 3L, 8L), stddev=0.1), 'w_conv1')b_conv1 <- tf$Variable(tf$truncated_normal(shape(1L, 8L), stddev=0.1), 'b_conv1') #w_conv2 <- tf$Variable(tf$truncated_normal(shape(4L, 4L, 8L, 8L), stddev=0.1), 'w_conv2')#b_conv2 <- tf$Variable(tf$truncated_normal(shape(1L, 8L), stddev=0.1), 'b_conv2') #w_conv3 <- tf$Variable(tf$truncated_normal(shape(4L, 4L, 8L, 8L), stddev=0.1), 'w_conv3')#b_conv3 <- tf$Variable(tf$truncated_normal(shape(1L, 8L), stddev=0.1), 'b_conv3') w_out <- tf$Variable(tf$truncated_normal(shape(1L, 1L, 8L, clas), stddev=0.1), 'w_upconv1')b_out <- tf$Variable(tf$truncated_normal(shape(1L, clas), stddev=0.1), 'b_upconv1') ####################################### De graphh_conv1 <- tf$nn$relu( tf$nn$conv2d(x, w_conv1, strides=c(1L, 1L, 1L, 1L), padding='SAME') + b_conv1 )#h_conv2 = tf$nn$relu( tf$nn$conv2d(h_conv1, w_conv2, strides=c(1L, 1L, 1L, 1L), padding='SAME') + b_conv2 )#h_conv3 = tf$nn$relu( tf$nn$conv2d(h_conv2, w_conv3, strides=c(1L, 1L, 1L, 1L), padding='SAME') + b_conv3 )out = tf$nn$conv2d(h_conv1, w_out, strides=c(1L, 1L, 1L, 1L), padding='SAME') + b_out out_softmax = tf$nn$softmax(out, dim = -1L)#################################fout functielabel_onehot = tf$one_hot(label, depth = clas)cost = - tf$reduce_mean( tf$multiply( label_onehot , tf$log( tf$clip_by_value(out_softmax, clip_value_min = 1e-10, clip_value_max = 1) ) ) )train_step <- tf$train$AdamOptimizer(lr)$minimize(cost)#################################maak sessiesess <- tf$InteractiveSession()sess$run(tf$global_variables_initializer())###############Now let's train this network.Now let's train this network.
x
for (i in 1:100) { #lees andom plaatjes in samp = sample( x= c(1: nrow(table)) , size = 1 ) print( as.character(table$naam[samp])) im = readImage( as.character(table$naam[samp]) ) im = array( im, dim = c(1,dim(im)) ) lab = array(0, dim = dim(im)[1:3]) lab[1,table$xmin[samp]:table$xmax[samp], table$ymin[samp]:table$ymax[samp]] = 1 #train met gradient descent train_step$run(feed_dict = dict(x = im , label = lab)) }Let's see how the network preforms on our test data.Let's see how the network preforms on our test data.
photos = list.files('db/test', full.name = TRUE)for(im in photos){im = readImage( as.character(table$naam[samp]) )im = array( im, dim = c(1,dim(im)) ) output = sess$run(out_softmax, feed_dict = dict(x = im))display(im[1,,,1])display(output[1,,,2]) }
If evrything went well the result should look a bit like the following images.<img src="notebook_images/3foto.jpg"><img src="notebook_images/3uitkomst.jpg">If evrything went well the result should look a bit like the following images.
In the previous step we obtained a heatmap in which pixels are brither when the computer deems it likely that the pixel is part of the buoy.If the sum of all pixels around a pixel cross a certain threshold we say it is part of a buoy. Next we downsize the photo a factor 10 to merge all piLet'sIn the previous step we obtained a heatmap in which pixels are brither when the computer deems it likely that the pixel is part of the buoy.
If the sum of all pixels around a pixel cross a certain threshold we say it is part of a buoy. Next we downsize the photo a factor 10 to merge all pi
Let's
photos = list.files('db/test', full.name = TRUE)thershold = 15sd_thershold1 = 100000positions = list()for( i in 1:nrow(locations)){ im = photos[i] im = readImage( im ) im = array( im, dim = c(1,dim(im)) ) output = sess$run(out_softmax, feed_dict = dict(x = im)) output = as.matrix(output[1,,,2])w = matrix(1, ncol = 11 , nrow = 11 )output_filter = filter2(output, w)output_filter = output_filter > thersholdposition_buoy = which(output_filter == max(output_filter), arr.ind = TRUE) #if sd is smaller than threshold_sd it is only one buoy if( sd(position_buoy[,1])*sd(position_buoy[,2]) < sd_thershold1 ){ k=1} if( sd(position_buoy[,1])*sd(position_buoy[,2]) > sd_thershold1 ){ k=2} new = as.data.frame( kmeans( position_buoy,k)$centers ) colnames(new)= c('col', 'row') positions[[i]] = new }print(positions)Based on the position of the buoy in the photo we can calculate the angle the buoy makes with the camera.The distance between the film and the lens was 4 cm and the width of the film was 5 cm.Based on the position of the buoy in the photo we can calculate the angle the buoy makes with the camera.
The distance between the film and the lens was 4 cm and the width of the film was 5 cm.
<img src="notebook_images/image.gif">
x
for( i in 1:length(positions)){ position_on_film = 5 * (positions[[i]]$col / 1296 - 0.5 ) positions[[i]]$angle = atan( position_on_film / 5) }print(positions)Based on the locations table constructed in the above we can calculate the position of the buoy.First we calculate the angle the boat makes with respect to the coordinate system. We add this value to the angle between the buoy and the camera with respect to the coordinate system. We add 0.25 * pi since the camera films perpendicular to the ships movement.Based on the locations table constructed in the above we can calculate the position of the buoy.
First we calculate the angle the boat makes with respect to the coordinate system. We add this value to the angle between the buoy and the camera with respect to the coordinate system. We add 0.25 * pi since the camera films perpendicular to the ships movement.
locations_next = rbind(locations, NA)[-1,]locations_rel = locations_next - locationslocations$angle_boat = tan( locations_rel$lon / locations_rel$lat )for( i in 1:length(positions)){positions[[i]]$absolute_angle = locations$angle_boat[i] + 0.25 * pi - positions[[i]]$angle }print(positions)Based on this we can calculate the coefficiet and offset of the line with respect to the coordinate system.Based on this we can calculate the coefficiet and offset of the line with respect to the coordinate system.
for( i in 1:length(positions)){ positions[[i]]$coeficient = tan(positions[[i]]$absolute_angle) positions[[i]]$offset = locations$lon[i] - positions[[i]]$coeficient * locations$lat[i] }print(positions)xxxxxxxxxxThe buoy lies on the intersection of these lines. We calculate the intersection of the first two photos.The buoy lies on the intersection of these lines. We calculate the intersection of the first two photos.
n = 1m= 2x_intersect = (positions[[n]]$offset[1] - positions[[m]]$offset[1]) / (positions[[m]]$coeficient[1] - positions[[n]]$coeficient[1])y_intersect = positions[[n]]$coeficient[1] * x_intersect + positions[[n]]$offset[1] print( paste('x_intersect:', x_intersect , 'y_intersect:' , y_intersect))We calculate the position in wgs coordinates and plot it in a leaflet.We calculate the position in wgs coordinates and plot it in a leaflet.
coords = c(x_intersect, y_intersect)python.assign('RDCoords', coords)python.exec('coords = conv.fromRdToWgs( RDCoords )')coords = python.get('coords')coordsm = leaflet()m = addTiles(m)m = addMarkers(m, lng= coords[2], lat= coords[1])m = addProviderTiles(m,providers$OpenStreetMap)saveWidget(m, file="buoy_location.html")